platform/dbops/binaries/weaviate-src/cluster/distributedtask/scheduler.go

//                           _       _
// __      _____  __ ___   ___  __ _| |_ ___
// \ \ /\ / / _ \/ _` \ \ / / |/ _` | __/ _ \
//  \ V  V /  __/ (_| |\ V /| | (_| | ||  __/
//   \_/\_/ \___|\__,_| \_/ |_|\__,_|\__\___|
//
//  Copyright © 2016 - 2025 Weaviate B.V. All rights reserved.
//
//  CONTACT: hello@weaviate.io
//

package distributedtask

import (
	"context"
	"fmt"
	"sync"
	"time"

	"github.com/jonboulle/clockwork"
	"github.com/prometheus/client_golang/prometheus"
	"github.com/prometheus/client_golang/prometheus/promauto"
	"github.com/sirupsen/logrus"
	enterrors "github.com/weaviate/weaviate/entities/errors"
	"github.com/weaviate/weaviate/usecases/logrusext"
	"github.com/weaviate/weaviate/usecases/monitoring"
)

// Scheduler is the component which is responsible for polling the active tasks in the cluster (via the Manager)
// and making sure that the tasks are running on the local node.
//
// The general flow of a distributed task is as follows:
// 1. A Provider is registered with the Scheduler at startup to handle all tasks under a specific namespace.
// 2. A task is created and added to the cluster via the Manager.AddTask.
// 3. Scheduler regularly scans all available tasks in the cluster, picks up new ones and instructs the Provider to execute them locally.
// 4. A task is responsible for updating its status in the cluster via TaskCompletionRecorder.
// 6. Scheduler polls the cluster for the task status and checks if it is still running. It cancels the local task if it is not marked as STARTED anymore.
// 7. After completed task TTL has passed, the Scheduler issues the Manager.CleanUpDistributedTask request to remove the task from the cluster list.
// 8. After a task is removed from the cluster list, the Scheduler instructs the Provider to clean up the local task state.
type Scheduler struct {
	mu           sync.Mutex
	runningTasks map[string]map[TaskDescriptor]TaskHandle

	providers          map[string]Provider // namespace -> Provider
	completionRecorder TaskCompletionRecorder
	tasksLister        TasksLister
	taskCleaner        TaskCleaner
	clock              clockwork.Clock

	localNode        string
	completedTaskTTL time.Duration
	tickInterval     time.Duration

	logger        logrus.FieldLogger
	sampledLogger *logrusext.Sampler

	tasksRunning *prometheus.GaugeVec

	stopCh chan struct{}
}

type SchedulerParams struct {
	CompletionRecorder TaskCompletionRecorder
	TasksLister        TasksLister
	TaskCleaner        TaskCleaner
	Providers          map[string]Provider
	Clock              clockwork.Clock
	Logger             logrus.FieldLogger
	MetricsRegisterer  prometheus.Registerer

	LocalNode        string
	CompletedTaskTTL time.Duration
	TickInterval     time.Duration
}

func NewScheduler(params SchedulerParams) *Scheduler {
	if params.Clock == nil {
		params.Clock = clockwork.NewRealClock()
	}

	if params.MetricsRegisterer == nil {
		params.MetricsRegisterer = monitoring.NoopRegisterer
	}

	return &Scheduler{
		runningTasks: map[string]map[TaskDescriptor]TaskHandle{},

		providers:          params.Providers,
		completionRecorder: params.CompletionRecorder,
		tasksLister:        params.TasksLister,
		taskCleaner:        params.TaskCleaner,
		clock:              params.Clock,

		localNode:        params.LocalNode,
		completedTaskTTL: params.CompletedTaskTTL,
		tickInterval:     params.TickInterval,

		logger:        params.Logger,
		sampledLogger: logrusext.NewSampler(params.Logger, 5, 5*params.TickInterval),

		tasksRunning: promauto.With(params.MetricsRegisterer).NewGaugeVec(prometheus.GaugeOpts{
			Name: "weaviate_distributed_tasks_running",
			Help: "Number of active distributed tasks running per namespace",
		}, []string{"namespace"}),

		stopCh: make(chan struct{}),
	}
}

func (s *Scheduler) Start(ctx context.Context) error {
	tasksByNamespace, err := s.listTasks(ctx)
	if err != nil {
		return fmt.Errorf("list distributed tasks: %w", err)
	}

	s.mu.Lock()
	defer s.mu.Unlock()
	for namespace, provider := range s.providers {
		provider.SetCompletionRecorder(s.completionRecorder)

		var (
			tasks         = tasksByNamespace[namespace]
			startedTasks  = s.filterStartedTasks(tasks)
			localTaskDesc = provider.GetLocalTasks()
		)
		for _, taskDesc := range localTaskDesc {
			if _, ok := startedTasks[taskDesc]; ok {
				continue
			}

			if err = provider.CleanupTask(taskDesc); err != nil {
				s.loggerWithTask(namespace, taskDesc).WithError(err).
					Error("failed to clean up local distributed task state")
				continue
			}

			s.loggerWithTask(namespace, taskDesc).Info("cleaned up local distributed task state")
		}

		for desc, task := range startedTasks {
			handle, err := provider.StartTask(task)
			if err != nil {
				return fmt.Errorf("provider %s start task %v: %w", namespace, desc, err)
			}

			s.setRunningTaskHandleWithLock(namespace, desc, handle)
			s.loggerWithTask(namespace, desc).Info("started distributed task execution")
		}

		s.tasksRunning.
			WithLabelValues(namespace).
			Set(float64(len(startedTasks)))
	}

	enterrors.GoWrapper(s.loop, s.logger)

	return nil
}

func (s *Scheduler) filterStartedTasks(tasks map[TaskDescriptor]*Task) map[TaskDescriptor]*Task {
	return filterTasks(tasks, func(task *Task) bool {
		return task.Status == TaskStatusStarted && !task.FinishedNodes[s.localNode]
	})
}

func filterTasks(tasks map[TaskDescriptor]*Task, predicate func(task *Task) bool) map[TaskDescriptor]*Task {
	filtered := make(map[TaskDescriptor]*Task, len(tasks))
	for _, task := range tasks {
		if !predicate(task) {
			continue
		}

		filtered[TaskDescriptor{
			ID:      task.ID,
			Version: task.Version,
		}] = task
	}
	return filtered
}

func (s *Scheduler) loop() {
	ticker := s.clock.NewTicker(s.tickInterval)
	defer ticker.Stop()

	for {
		select {
		case <-ticker.Chan():
			s.tick()
		case <-s.stopCh:
			return
		}
	}
}

func (s *Scheduler) tick() {
	tasksByNamespace, err := s.listTasks(context.Background())
	if err != nil {
		s.sampledLogger.WithSampling(func(l logrus.FieldLogger) {
			l.WithError(err).Error("failed to list distributed tasks")
		})
		return
	}

	s.mu.Lock()
	defer s.mu.Unlock()

	for namespace, provider := range s.providers {
		tasks := tasksByNamespace[namespace]

		// Check that all tasks that are supposed to be running
		// and launch if they aren't.
		startedTasks := s.filterStartedTasks(tasks)
		for _, activeTask := range startedTasks {
			if _, alreadyLaunched := s.runningTasks[namespace][activeTask.TaskDescriptor]; alreadyLaunched {
				continue
			}

			handle, err := provider.StartTask(activeTask)
			if err != nil {
				s.sampledLogger.WithSampling(func(l logrus.FieldLogger) {
					s.loggerWithTask(namespace, activeTask.TaskDescriptor).WithError(err).
						Error("failed to start distributed task")
				})
				continue
			}

			s.setRunningTaskHandleWithLock(namespace, activeTask.TaskDescriptor, handle)
			s.loggerWithTask(namespace, activeTask.TaskDescriptor).Info("started distributed task execution")
		}

		s.tasksRunning.
			WithLabelValues(namespace).
			Set(float64(len(startedTasks)))

		// Check that all tasks that are not supposed to be running are not running.
		for desc, taskHandle := range s.runningTasks[namespace] {
			if _, ok := startedTasks[desc]; ok {
				continue
			}

			taskHandle.Terminate()
			delete(s.runningTasks[namespace], desc)

			s.loggerWithTask(namespace, desc).Info("terminated distributed task execution")

		}

		// Check that all tasks that are already finished and if their TTL has passed, so we can clean them up.
		cleanableTasks := filterTasks(tasks, func(task *Task) bool {
			return task.Status != TaskStatusStarted && s.completedTaskTTL <= s.clock.Since(task.FinishedAt)
		})
		for _, task := range cleanableTasks {
			err = s.taskCleaner.CleanUpDistributedTask(context.Background(), namespace, task.ID, task.Version)
			if err != nil {
				s.sampledLogger.WithSampling(func(l logrus.FieldLogger) {
					s.loggerWithTask(namespace, task.TaskDescriptor).WithError(err).
						Error("failed to clean up distributed task")
				})
				continue
			}

			s.loggerWithTask(namespace, task.TaskDescriptor).
				Info("successfully submitted request to clean up distributed task")
		}

		// Check that tasks that can be cleaned up locally
		localTasks := provider.GetLocalTasks()
		for _, desc := range localTasks {
			if _, ok := tasks[desc]; ok {
				// task still present in the list
				continue
			}

			if err = provider.CleanupTask(desc); err != nil {
				s.sampledLogger.WithSampling(func(l logrus.FieldLogger) {
					s.loggerWithTask(namespace, desc).WithError(err).
						Error("failed to clean up local distributed task state")
				})
			}
		}
	}
}

func (s *Scheduler) listTasks(ctx context.Context) (map[string]map[TaskDescriptor]*Task, error) {
	tasksByNamespace, err := s.tasksLister.ListDistributedTasks(ctx)
	if err != nil {
		return nil, fmt.Errorf("list distributed tasks: %w", err)
	}

	result := make(map[string]map[TaskDescriptor]*Task, len(tasksByNamespace))
	for namespace, tasks := range tasksByNamespace {
		result[namespace] = make(map[TaskDescriptor]*Task, len(tasks))
		for _, task := range tasks {
			result[namespace][task.TaskDescriptor] = task
		}
	}
	return result, nil
}

func (s *Scheduler) setRunningTaskHandleWithLock(namespace string, desc TaskDescriptor, handle TaskHandle) {
	if _, ok := s.runningTasks[namespace]; !ok {
		s.runningTasks[namespace] = map[TaskDescriptor]TaskHandle{}
	}
	s.runningTasks[namespace][desc] = handle
}

func (s *Scheduler) Close() {
	close(s.stopCh)

	s.mu.Lock()
	defer s.mu.Unlock()

	for _, tasks := range s.runningTasks {
		for _, task := range tasks {
			task.Terminate()
		}
	}
}

func (s *Scheduler) totalRunningTaskCount() int {
	s.mu.Lock()
	defer s.mu.Unlock()

	count := 0
	for _, tasks := range s.runningTasks {
		count += len(tasks)
	}
	return count
}

func (s *Scheduler) loggerWithTask(namespace string, taskDesc TaskDescriptor) *logrus.Entry {
	return s.logger.WithFields(logrus.Fields{
		"namespace":   namespace,
		"taskID":      taskDesc.ID,
		"taskVersion": taskDesc.Version,
	})
}